fix(sql): avoid excessive inlining during Select merge

[kszucs]

Another take on #8820 by introducing a complexity metric and only merge Select nodes if the outcome is less complex.

[cpcloud]

@NickCrews If you want to take this for a spin that would be great.

After this PR, the number of lines of SQL for 8.0.0 and in this PR are at the same order of magnitude, which IMO is enough of an improvement to merge it.

Here's a file with each version's generated SQL from the notebook provided in #8484:

8.0.0-nb.txt
8825-nb.txt

(they're .txt because GH doesn't allow uploading .sql files for some reason)

[kszucs]

@cpcloud I added docstrings to the complexity calculation.

[cpcloud]

@kszucs Should this close #8484?

[NickCrews]

Thank you so much for this! I just tested with the notebook in #8484

• 8.0.0: fastest
• this PR: ~10% slower than 8.0.0
• main: 30+ times slower than 8.0.0

So this is definitely an improvement.

The theory of the implementation sounds right, even if the details go over my head a bit.

[cpcloud]

Ideally over time we can get back the 10%, but probably not before the 9.0 release.

I think this is the last major blocker for 9.0.